Gruow

Thinking

Less thinking, more checking: the new core skill is judgment

Judgment is the core skill of knowledge work now: knowing what good looks like, spotting what’s wrong, and knowing when it matters.

For most of working history, the hard part was producing the thing: the draft, the analysis, the code. AI now produces a decent first version of all three in seconds. The hard part has moved. It’s deciding whether that version is right.

That makes judgment the core skill of knowledge work: knowing what good looks like, spotting what’s wrong, and knowing when it matters. It’s also the skill that heavy AI use wears down fastest, unless people protect it on purpose.

None of this is new. People have worried about it for 2,400 years, and they were partly right.

We’ve worried about this before

In Plato’s Phaedrus, written around 370 BC, the god Theuth presents writing to the Egyptian king Thamus as a cure for forgetting. Thamus isn’t convinced. In Benjamin Jowett’s translation:

This discovery of yours will create forgetfulness in the learners’ souls, because they will not use their memories… they will be hearers of many things and will have learned nothing; they will appear to be omniscient and will generally know nothing; they will be tiresome company, having the show of wisdom without the reality.

Thamus was wrong that writing would end thinking. It moved thinking somewhere else. Once anyone could write things down, the valuable skill became reading well: knowing who wrote it, whether it holds up, and what it leaves out. But his last line still lands. “The show of wisdom without the reality” is a fair description of an AI answer that’s almost right.

In 1983, the engineering psychologist Lisanne Bainbridge made the same point about factories and power plants in a short paper called “Ironies of Automation”. When designers automate a process, she wrote, the human is left with two jobs: watching the system, and taking over when it fails. Those are the hardest jobs there are. People can’t keep close attention on a system where little happens for more than about half an hour. And the skills needed to take over fade from lack of use: “physical skills deteriorate when they are not used.”

Her central irony was that the more we automate, the more we depend on a skilled human to catch what the automation gets wrong, and the fewer chances that human gets to stay skilled. Swap the control room for a laptop and the paper reads like it was written this year.

Confident, and checking less

The research on AI at work keeps finding the same thing: the more people trust the tool, the less they check it.

Researchers from Microsoft and Carnegie Mellon surveyed 319 knowledge workers about 936 real tasks they’d done with AI. Those more confident in AI applied less critical thinking to its output. Those more confident in their own skills applied more. The thinking didn’t disappear; it changed jobs. Less went into finding information and solving problems from scratch, and more into verifying, adapting and overseeing what the AI produced. The catch, in the authors’ words: people “often refrain from critical thinking when they lack the skills to inspect, improve, and guide” what AI gives them.

At a larger scale, KPMG and the University of Melbourne surveyed more than 48,000 people in 47 countries. 66% said they rely on AI output without checking its accuracy, and 56% said they’ve made mistakes at work because of AI.

Developers are the furthest along, and they’re the most wary. In Stack Overflow’s 2025 survey of more than 49,000 developers, 84% use or plan to use AI tools, but only 33% trust their accuracy and 46% distrust it. Their top frustration, named by 66%, is answers that are “almost right, but not quite”.

Even our sense of whether AI is helping can be off. METR, a research group, ran a trial with 16 experienced open-source developers on 246 real tasks from their own projects.

Developers felt about 20% faster with AI. They were 19% slower.

Speed with AI against without it, 16 experienced developers on 246 real tasks, early 2025

Source: METR, July 2025. Redrawn by Gruow.

The developers weren’t careless. They were experts on their own code, and they still couldn’t tell from the inside that the tool was slowing them down. Judgment isn’t only about the AI’s output. It’s also about judging our own work with it.

When nobody checks

The cost of skipped checking is now public, and it lands on people paid to be careful.

In 2025, Deloitte delivered a AU$440,000 review of a welfare compliance system to Australia’s Department of Employment and Workplace Relations. A University of Sydney academic, Christopher Rudge, found it cited reports that don’t exist and quoted a federal court judgment that says no such thing. Deloitte later disclosed it had used a GPT-4o-based tool, corrected more than a dozen references and footnotes, and agreed to refund the final payment. The department said the recommendations didn’t change. The trust did.

Courts see the same thing weekly. A database kept by the researcher Damien Charlotin had logged 1,058 US court cases involving AI-made errors such as invented case law by June 2026. About 40% involved lawyers, not people representing themselves.

The quieter cost is Bainbridge’s second irony: skill fading with disuse. A study in The Lancet Gastroenterology & Hepatology followed 19 experienced endoscopists in Poland, each with more than 2,000 colonoscopies behind them. After AI-assisted colonoscopy became routine, their rate of finding precancerous growths in procedures done without AI fell from 28.4% to 22.4% within months. The tool helped while it was on. The doctors got worse at the job when it was off.

What good judgment looks like

Checking AI isn’t proofreading. Reading for typos won’t catch a quote that was never said. Judgment means three things: knowing what good looks like in your field, knowing where AI tends to be wrong, and knowing what’s at stake if it is.

A few habits follow from the evidence:

  • Check in proportion to the stakes. A social post and a report to a government department don’t need the same review. Decide which work gets a second look before anyone starts, not after something goes wrong.
  • Check the claims, not the prose. AI writing reads well. The errors hide in names, numbers, dates, quotes and citations, which is exactly where Deloitte and the lawyers went wrong. Open every source the work rests on.
  • Keep doing some of the work yourself. The endoscopists lost skill in months. If a task matters, do it without AI often enough to stay able to judge it with AI.
  • Measure, don’t trust the feeling. The METR developers felt faster and weren’t. Time the work, count the fixes, and compare.
  • Build expertise before handing over. The people who checked AI most in the Microsoft study were the ones most confident in their own skills. Judgment rests on knowing the work, which is why how new people learn it matters more now, not less.

In the systems we build at Gruow, an agent can draft, sort and suggest, and a person approves anything that matters. That isn’t caution for its own sake. It’s where the judgment lives.

Thamus was right that a new tool changes what thinking is for, and wrong that it ends it. The skill now is telling the show of wisdom from the real thing.

Sources